NequIP

Overview

NequIP is a software framework for building E(3)-equivariant interatomic potentials, rather than a single fixed predictive model. It supports a workflow spanning model training, compilation and use in atomistic simulations. The project README also describes pre-trained foundation potentials covering most of the periodic table, available through nequip.net for direct simulation use or adaptation to a user's own dataset.

Users can start with their own training data or a supplied pre-trained potential. The documented workflow produces trained or compiled potentials for downstream use, including through an ASE calculator and LAMMPS integrations. The excerpts establish these workflow roles but do not specify dataset schemas, required labels or exact prediction interfaces; those details need to be checked in the linked guides. Package entry points cover training, packaging, compilation and preparation for the LAMMPS ML-IAP integration.

For computational workflows, NequIP documents compiled training and inference, multi-GPU training, and GPU kernel acceleration through OpenEquivariance and CuEquivariance; the latter is marked alpha. Extension packages allow custom architectures and training methods, with Allegro identified as a separate extension implementing a strictly local equivariant architecture.

Existing configurations require attention to release compatibility: the README identifies v0.7.0 as a backwards-incompatible update and points users of older configurations to v0.6.2. The source excerpts contain no numerical accuracy or speed results, and they do not establish suitability for a particular chemical system. The repository's MIT code licence should not be assumed to define the terms of separately distributed model weights or datasets.

Key Features

• Framework for building E(3)-equivariant interatomic potentials, with extension support for custom architectures and training techniques.
• Pre-trained foundation potentials distributed through nequip.net, with documented workflows for downloading, compiling, running and fine-tuning them.
• Compiled training and inference workflows, plus multi-GPU training.
• GPU kernel acceleration integrations with OpenEquivariance and CuEquivariance, with CuEquivariance marked alpha in the README.
• ASE calculator integration and LAMMPS support through pair_nequip_allegro pair styles and an ML-IAP integration.

Use Cases

• Intended evaluation: assess a foundation potential on representative structures from a target materials system before using it in broader simulations.
• Intended evaluation: fine-tune a pre-trained potential on a project-specific dataset and assess its suitability for the intended chemical domain.
• Intended evaluation: connect a trained or compiled potential to an ASE or LAMMPS workflow and check the relevant integration requirements.
• Intended evaluation: explore a custom architecture or training method through an extension package, using Allegro as a documented example of this mechanism.

How to Use

1. Begin with the installation and user guides. Check the requirements for your chosen workflow; the supplied package metadata declares Python >=3.10, while the README warns that v0.7.0 introduced backwards-incompatible changes.
2. Work through the tutorial notebook. The README describes it as running entirely on Google Colab without local installation.
3. Choose between training on your own data and starting from a pre-trained potential. Consult the foundation potentials guide for download, compilation and execution instructions, including any applicable model terms.
4. If adapting a pre-trained model, follow the fine-tuning guide. As a suggested evaluation step, assess representative cases from your target system before expanding use.
5. Select the ASE integration or LAMMPS integration appropriate to your simulation workflow. Check compilation and integration instructions rather than assuming interchangeable deployment paths.